Diagnostic Uncertainty and Antibiotic Timing in Sepsis: Observational Validation of a Dual-Axis Decision Model.

IMPORTANCE: Sepsis treatment requires balancing the benefits of timely antibiotic administration against the risks of over-treatment under diagnostic uncertainty. Although treatment urgency likely reflects both the probability of infection and physiologic severity, this joint structure has not previously been quantified using real-world clinical data. OBJECTIVES: To determine whether real-world antibiotic timing in the emergency department reflects a dual-axis pattern integrating probability of infection and physiologic severity. DESIGN, SETTING, AND PARTICIPANTS: Retrospective observational cohort study using structured electronic health record data from adult emergency department encounters at a large urban safety-net health system from January 2019 through June 2025. A random forest model estimated the probability of sustained antibiotic treatment, P(treated infection), for suspected infection from early demographic, physiologic, and laboratory data, and physiologic severity was measured using the Modified Early Warning Score at 3 hours. MAIN OUTCOMES AND MEASURES: The primary outcome was time-to-antibiotics within 24 hours of presentation. The secondary outcome was a composite of ICU stay or in-hospital death among patients treated after 3 hours. RESULTS: The study included 199,062 emergency department visits over 6.5 years. Of these, 83,910 visits were used for model development, and the remaining 115,152 emergency department encounters comprised a temporally held-out cohort used for model evaluation and downstream observational analyses. The P(treated infection) model demonstrated good discrimination (C-statistic, 0.86). Antibiotic timing followed a continuous gradient across jointly increasing P(treated infection) and physiologic severity, and restricted cubic spline regression confirmed significant contributions from both axes and their interaction (interaction, p = 0.0013). Among patients treated after 3 hours, antibiotic timing later-than-expected based on a patient's dual-axis position was associated with a higher rate of 3-day ICU stay or in-hospital death after adjustment for P(treated infection), physiologic severity, and their interaction (p < 0.001). CONCLUSIONS: Real-world antibiotic timing in suspected sepsis appears to reflect a dual-axis pattern consistent with clinicians integrating P(treated infection) and physiologic severity. Making this structure explicit may help identify unwarranted practice variation and inform future decision support strategies for timely treatment.